Abstract: To address the performance optimization issues in federated learning for unmanned aerial vehicle (UAV) digital twin edge networks, a resource scheduling strategy is proposed based on deep reinforcement learning for UAV digital twin edge networks. Considering dynamic and time varying features of UAV digital twin edge networks environment, a twin network model is built consisting of base station (BS) , intelligent terminals, UAV and wireless transmission channel. Then an adaptive resource optimization model is formulated which jointly optimized UAV flying distance, flying angle and spectrum resource allocation, in order to minimize time delay of federated learning. Moreover, an UAV digital twin edge networks empowered multi-agent deep deterministic policy gradient (MA-DDPG) algorithm is designed to solve the adaptive resource optimization model. The algorithm training process adopts centralized training and decentralized execution. Each UAV agent considers the states and actions of other agents when evaluating the value of actions, but decides actions based only on its own local observations during execution. The above training process is conducted in digital twin environment, and after the algorithm converges, and it is applied to the real world, minimizing the resource overhead of physical entities to the greatest extent. Numerical results illustrate that the proposed algorithm can significantly decrease the service latency of federated learning, while guaranteeing the superiority of training loss and accuracy of federated learning.

Key words: unmanned aerial vehicle networks, digital twin, federated learning, multi-agent deep deterministic policy gradient